| Presentation | 2022-09-29 Isotropic Inference and Modularity in Neural Networks Yuma Onishi, Tomoaki Imajo, Yusuke Matsubara, Koiti Hasida, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Neural networks can be regarded as a system of simultaneous equations. Reasoning with it is to solve the simultaneous equations (constraint satisfaction). Considering it as a constraint satisfaction problem, multiple neural networks can be linked by sharing units (variables). In order for the simultaneous equations to be solved, the variables and the number of equations must be aligned. This can be done by deriving the equations of the neural network from the energy minimization principle, in which the energy function should be a loss function. Then, to minimize that energy function, a numerical solution method is used. State-of-the-art results in inference and learning using numerical solution methods have been achieved, for example, in the Deep Equilibrium Model. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | energy function / energy minimization principle / free energy principle / isotropic inference |
| Paper # | NC2022-37 |
| Date of Issue | 2022-09-22 (NC) |
| Conference Information | |
| Committee | NC / MBE |
|---|---|
| Conference Date | 2022/9/29(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Tohoku Univ. |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | Brain Architecture, NC, ME |
| Chair | Hiroshi Yamakawa(Univ of Tokyo) / Junichi Hori(Niigata Univ.) |
| Vice Chair | Hirokazu Tanaka(Tokyo City Univ.) / Hisashi Yoshida(Kinki Univ.) |
| Secretary | Hirokazu Tanaka(NTT) / Hisashi Yoshida(NICT) |
| Assistant | Yoshimasa Tawatsuji(Waseda Univ.) / Tomoki Kurikawa(KMU) / Emi Yuda(Tohoku Univ) / Miki Kaneko(Osaka Univ.) |
| Paper Information | |
| Registration To | Technical Committee on Neurocomputing / Technical Committee on ME and Bio Cybernetics |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Isotropic Inference and Modularity in Neural Networks |
| Sub Title (in English) | |
| Keyword(1) | energy function |
| Keyword(2) | energy minimization principle |
| Keyword(3) | free energy principle |
| Keyword(4) | isotropic inference |
| 1st Author's Name | Yuma Onishi |
| 1st Author's Affiliation | The University of Tokyo(UTokyo) |
| 2nd Author's Name | Tomoaki Imajo |
| 2nd Author's Affiliation | The University of Tokyo(UTokyo) |
| 3rd Author's Name | Yusuke Matsubara |
| 3rd Author's Affiliation | The University of Tokyo(UTokyo) |
| 4th Author's Name | Koiti Hasida |
| 4th Author's Affiliation | The University of Tokyo(UTokyo) |
| Date | 2022-09-29 |
| Paper # | NC2022-37 |
| Volume (vol) | vol.122 |
| Number (no) | NC-195 |
| Page | pp.pp.20-23(NC), |
| #Pages | 4 |
| Date of Issue | 2022-09-22 (NC) |